SAIG

Prepare

cd ./pytorch-image-models
pip install -e .

Please download CVUSA, CVACT, VIGOR. Please download the pre-trained model

Dataset	Model	top-1	weight	log
ImageNet-1k	SAIG-S	77.2%	weight	log
ImageNet-1k	SAIG-D	80.3%	weight	log

CVUSA & CVACT

bash train.sh
bash train_sam.sh

or directly run python train.py or python train_sam.py by giving other args

VIGOR

bash train_vigor.sh
bash train_vigor_sam.sh

CVUSA & CVACT

bash test.sh

VIGOR

bash test_vigor.sh

CVUSA

Method	Pool	Loss	ASAM	R@1	Weight
SAIG-S	GAP	Triplet	No	88.82	Google_Drive
SAIG-S	GAP	Triplet	Yes	92.69	Google_Drive
SAIG-S	SMD	Triplet	No	91.77	Google_Drive
SAIG-S	SMD	Triplet	Yes	95.37	Google_Drive
SAIG-D	GAP	Triplet	No	90.29	Google_Drive
SAIG-D	GAP	Triplet	Yes	93.97	Google_Drive
SAIG-D	SMD	Triplet	No	92.71	Google_Drive
SAIG-D	SMD	Triplet	Yes	96.08	Google_Drive

CVACT

Method	Pool	Loss	ASAM	R@1	Weight
SAIG-S	GAP	Triplet	No	81.39	Google_Drive
SAIG-S	GAP	Triplet	Yes	85.39	Google_Drive
SAIG-S	SMD	Triplet	No	83.54	Google_Drive
SAIG-S	SMD	Triplet	Yes	88.44	Google_Drive
SAIG-D	GAP	Triplet	No	82.40	Google_Drive
SAIG-D	GAP	Triplet	Yes	86.65	Google_Drive
SAIG-D	SMD	Triplet	No	84.42	Google_Drive
SAIG-D	SMD	Triplet	Yes	89.21	Google_Drive

VIGOR

Method	Pool	Loss	ASAM	Same Area R@1	Cross Area R@1
SAIG-S	GAP	Triplet	No	40.38	10.22
SAIG-S	GAP	Triplet	Yes	46.21	15.33
SAIG-D	GAP	Triplet	No	42.15	11.88
SAIG-S	SMD	Triplet	No	45.92	14.50
SAIG-D	SMD	Triplet	No	51.50	17.58
SAIG-D	SMD	InfoNCE	No	55.37	23.47
SAIG-D	SMD	Semi-Triplet	No	55.60	22.35
SAIG-S	SMD	Triplet	Yes	57.57	25.44
SAIG-S	SMD	Semi-Triplet	Yes	62.28	30.14
SAIG-D	SMD	Triplet	Yes	61.27	27.61
SAIG-D	SMD	InfoNCE	Yes	62.92	32.77
SAIG-D	SMD	Semi-Triplet	Yes	65.23	33.05